Posters

Abstract

By mid-2025, the global wind energy sector hit another milestone, reaching 1.27 TW. This expansion, particularly into remote and complex environments, has amplified operational challenges. In the pursuit of sustaining 95% availability, significant resources—between 20% and 25% of the levelized cost of energy—have been absorbed by operation and maintenance (O&M), raising concerns about projects' competitiveness and complexities. In this context, conventional maintenance strategies falter: preventive approaches yield inefficiencies with unnecessary interventions, while reactive models lag in addressing failures. Data scarcity, sensor constraints, and coordination demands lead maintenance crews into sub-optimal policies and actions. As a result, suboptimal operations lead to recurrent downtime and inflated costs, stemming from models ill-suited to partial observability and uncertainty. To overcome these barriers, research has advanced intelligent, condition-based maintenance frameworks that leverage artificial intelligence through pipelines. One of them is reinforcement learning (RL), a paradigm that optimises decisions through trial-and-error and reward signals that has recently proven effective in wind power applications. Building on this trajectory, the present work proposes multi-agent reinforcement learning (MARL) architectures as a means of bridging the gap between data uncertainties and management decisions. This proposal features a novel pipeline feeding a council of specialized agents under partial observability, with component-level agents monitoring turbine states and cluster-level supervisors coordinating actions and final policies. Such a configuration creates a decentralized yet orchestrated maintenance ecosystem theoretically capable of adapting to dynamic conditions, overcoming incomplete information, and supporting resilient decisions across heterogeneous data sources. Ultimately, MARL’s interactive loop feeds from downtime records, fault patterns, and anomaly detection for continuous learning, adapting to aging equipment and crew needs. Thus, by transforming raw operational data into coordinated maintenance actions, this framework may not only enhance turbine operation but also foster a pipeline for next-generation smart self-optimizing management systems for uncertain and dynamic environments.